World models must learn when to stop imagining
Peking University-led review defines the "computational usability gap" for robots and autonomous driving: AI world models must decide what to compute, how far to simulate, where to run it, and whether it pays off
image:
Computational usability asks what to compute, how far to simulate, where to run it, and what value it adds.
Action-oriented data help models prioritize the states that matter for decisions and safety.
Risk-adaptive rollouts and device-edge-cloud computing may make world models more deployable.
Action gain per computational cost measures improvements in decision-making relative to computing and resource costs.
view more
Credit: Song Y., Wang J., Liu H., et al. (2026). World models: Toward computationally usable embodied AI. The Innovation Insights 1:100018.
The most useful world model for a robot may not be the one that generates the most realistic video, but the one that knows when to stop imagining. A review published October 20 in The Innovation Insights argues that the competitive focus of world models — AI systems that predict how environments evolve — is shifting from visual fidelity to "computational usability": whether each unit of computation actually improves decisions enough to justify its cost. World models have expanded rapidly, from agents training policies in self-generated "dreams" to video generators and interactive 3D worlds. Meta's V-JEPA 2 learns from more than one million hours of internet video; DeepMind's Genie 3 generates 720p interactive worlds at 24 frames per second; NVIDIA's Cosmos packages world foundation models into a physical-AI platform. But robots, vehicles, and industrial systems cannot generate futures offline with unlimited computing. They must act within milliseconds to seconds under partial observations, limited energy and bandwidth, and real-time safety requirements. The review, led by Jianxiao Wang of Peking University with colleagues at the University of Hong Kong, Sun Yat-sen University, and JD's Joy Future Academy, calls this mismatch the "computational usability gap." The authors organize it as four coupled questions: what to compute, how far to roll out a simulation, where to compute, and whether the result is worth its cost. Evidence is mounting that more compute does not buy more usefulness. On the same consumer GPU, a latent-rollout planner (V-JEPA 2-AC) cut planning latency from 240 to 16 seconds per action compared with the pixel-generative Cosmos-7B, while raising reach success from 80% to 100%. A communication-aware method (ComVLA) shrank transmitted data from 1,088 to 68 kilobytes per step and computation from 8.8 to 2.3 teraflops at a cost of only 1.5 percentage points in task success. Realism, meanwhile, does not guarantee actionability: under the MiraBench evaluation, all 12 tested world-model configurations scored below 26 on physical-rule adherence even though 5 of 12 exceeded 90 on task completion. The authors propose a unifying metric, action gain per computational cost (AGC): the improvement a model delivers over a baseline, divided by weighted computation, energy, latency, and communication costs. As a proof of concept on matched real-robot results, the efficient Efficient-WAM model scored roughly 40 times higher on AGC than the stronger baseline Motus. The authors caution that AGC is a proof of concept from a single comparison, and note that energy consumption is rarely reported alongside task performance in current studies. They call for a hierarchical evaluation loop spanning visual and physical checks offline, counterfactual rollouts in simulation, task success and latency on real machines, and communication overhead and safety redundancy across device-edge-cloud systems. Large-scale generative models will remain an important foundation, the authors conclude. But physical AI requires not unconstrained imagination, but imagination that is schedulable, verifiable, and deployable.